| Panax notoginseng is one of the Chinese herbal medicines with Chinese characteristics.With the upgrading of modern fresh panax notoginseng sorting industry,the traditional manual cutting after the taproot and rhizome of manual sorting into machine cutting after manual sorting.The cost and efficiency of manual sorting and the difficulty of machine sorting have seriously affected the development of fresh Panax notoginseng mechanized industry.An automatic sorting device for fresh panax notoginseng taproot and rhizome was urgently needed in the market.Therefore,the rapid sorting of taproot and rhizome of Panax notoginseng was studied systematically based on machine vision technology.The main research contents and conclusions are as follows:1.This study established a single objective sorting model of Panax Notoginseng based on improved YOLOv4-Tiny network.In this model,Efficient Channel Attention(ECA)module was introduced into Feature Pyramid Network(FPN)structure of YOLOv4-Tiny network,Pyconv module was added into neck network,Complete Intersection over Union(CIo U)was used as loss function to optimize anchor frame and other operations to improve model detection accuracy and efficiency.The precision changes of fresh Panax Notoginseng cleaning factors before and after model improvement were explored through experiments,and the modeling effects of different attention mechanisms and different typical models were compared.The results showed that all kinds of image features of fresh Panax Notoginseng before cleaning were more obvious,and the modeling effect was better.The Mean Average Precision(m AP)value was increased by 5.45% to 96.34% compared with that after cleaning.The m AP value of the improved YOLOv4-Tiny model on the single target dataset of fresh Panax Notoginseng before cleaning was increased by 1.16% to 97.50%.ECA was the best in the comparison of the four attention mechanisms,with m AP value increased by 0.32%to 96.66% compared with the original model.The effectiveness of each improved module was verified by the ablation test between the improved models.Compared with different typical models,the I-YOLOv4-Tiny model had the best comprehensive test effect.It was the most suitable model for fresh Panax Notoginseng multi-objective sorting equipment.The pre-training model and research method were provided for the subsequent multi-target separation of fresh Panax Notoginseng.2.Determine the design requirements and technical parameters of the fresh Panax Notoginseng sorting system according to its working principle.Select specifications and models for cameras,high-power power supply groups,large-capacity air compressors,industrial computer and other equipment.Design the system hardware sorting structure.Draw the electrical schematic diagram of the system hardware platform.The function of the selection and installation position of each device was analyzed and expounded.In order to solve the problem of material repeated identification in visual field during video stream processing,this study designed and compared two software control strategies.Automatic sorting of fresh Panax Notoginseng was realized by writing system control software linkage system hardware.Design man-machine interaction interface to display and set the operating parameters and real-time images of the system.Encapsulate the software to solve the problem of complex operating environment configuration in the process of software migration,and increase the portability and expansibility of the system.3.The multi-target data set of fresh Panax Notoginseng was collected and produced by means of video streaming.The improved YOLOv4-Tiny model of Fresh Panax Notoginseng single object data set training was transferred to multi-object sorting.The influence of transfer learning on model training was explored.The modeling effect of different typical models on multi-objective separation of Panax notoginseng was compared and analyzed.The results show that the m AP value of the migrated model is increased by 1.59% compared with the original model.Compared with the original model,the m AP value of the improved YOLOv4-Tiny model was increased by 2.5% to 89.23%.Compared with the modeling effect of other typical models,the I-YOLOv4-Tiny model can achieve the optimal detection speed and model size under the condition that the model accuracy is better.The optimum multi-objective separation model of Panax Notoginseng was provided for the test system.4.According to the test requirements of the system,the evaluation indexes of the performance and reliability of the reaction system were designed,such as the theoretical sorting speed of the equipment,the error output rate of the equipment,the error rate of the system and the sorting rate.Combined with the hardware conditions of the system,the condition table of the system image acquisition without dragging image,the theoretical calculation and measured value of the conveyor belt surface velocity,and the calculation formula of the accurate delay positioning time of materials were obtained by calculating the parameters of the equipment.The reliable technical parameter reference was provided for the multi-objective separation test of fresh Panax Notoginseng.5.In this paper,two different software control strategies were used to solve the problem of multiple identification and repeated output in the field of visual field during real-time sorting of materials.The first control strategy was to make each material appear once in the field of vision by controlling the speed of conveyor belt and the frame rate of camera acquisition.In the second control strategy,Simple Online and Realtime Tracking(SORT)algorithm is adopted to track the materials within the scope of visual field and remove the duplicate results in the tracking process.The experimental results of two control strategies under different feeding densities and different speeds were compared and analyzed.The results show that when the conveyor belt speed was 330mm/s and the feeding mode is low density,the sorting accuracy of control strategy 1 reaches the highest,which was 93.5%.At the same time,the maximum separation efficiency of the system was 580.2kg/h when the separation accuracy was above 90%.When the conveyor belt speed was 150mm/s and the feeding mode was low density,the sorting accuracy of control strategy 2 was as high as 90%,and the sorting efficiency was as high as 150.7kg/h.Although control strategy one cannot completely solve the problem of multiple repetition identification,its test results were optimal and it was the current optimal control method for equipment.The sorting speed of control strategy two was slow and improves the identification error rate of single material,but greatly reduces the power demand pressure and the maintenance pressure of air source pressure during the operation of the equipment,which was conducive to the later improvement of the equipment.In short,this paper set up a set of fresh panax Notoginseng rapid sorting system based on machine vision technology.Multiple batches of fresh panax notoginseng raw materials from multiple origins were used.The universality and extensibility of the sorting model and equipment in this study were verified.The separation tests of two control strategies were designed and analyzed.It laid a theoretical and technical foundation for improving the accuracy and efficiency of real-time separation of Panax Notoginseng in the future. |
PDF Full Text Request